This attribute will cause us to invoke evaluate on every where clause of an
invoked function and to generate an error with the result.
Without this, it is very difficult to observe the effects of invoking the trait
evaluator.
Emit errors/warns on some wrong uses of rustdoc attributes
This PR adds a few diagnostics:
- error if conflicting `#[doc(inline)]`/`#[doc(no_inline)]` are found
- introduce the `invalid_doc_attributes` lint (warn-by-default) which triggers:
- if a crate-level attribute is used on a non-`crate` item
- if `#[doc(inline)]`/`#[doc(no_inline)]` is used on a non-`use` item
The code could probably be improved but I wanted to get feedback first. Also, some of those changes could be considered breaking changes, so I don't know what the procedure would be? ~~And finally, for the warnings, they are currently hard warnings, maybe it would be better to introduce a lint?~~ (EDIT: introduced the `invalid_doc_attributes` lint)
Closes#80275.
r? `@jyn514`
Fix stack overflow when checking for structural recursion
This pull request aims to fix#74224 and fix#84611. The current logic for detecting ADTs with structural recursion is flawed because it only looks at the root type, and then for exact matches. What I mean by this is that for examples such as:
```rust
struct A<T> {
x: T,
y: A<A<T>>,
}
struct B {
z: A<usize>
}
fn main() {}
```
When checking `A`, the compiler correctly determines that it has an infinite size (because the "root" type is `A`, and `A` occurs, albeit with different type arguments, as a nested type in `A`).
However, when checking `B`, it also recurses into `A`, but now `B` is the root type, and it only checks for _exact_ matches of `A`, but since `A` never precisely contains itself (only `A<A<T>>`, `A<A<A<T>>>`, etc.), an endless recursion ensues until the stack overflows.
In this PR, I have attempted to fix this behavior by implementing a two-phase checking: When checking `B`, my code first checks `A` _separately_ and stops if `A` already turns out to be infinite. If not (such as for `Option<T>`), the second phase checks whether the root type (`B`) is ever nested inside itself, e.g.:
```rust
struct Foo { x: Option<Option<Foo>> }
```
Special care needs to be taken for mutually recursive types, e.g.:
```rust
struct A<T> {
z: T,
x: B<T>,
}
struct B<T> {
y: A<T>
}
```
Here, both `A` and `B` both _are_ `SelfRecursive` and _contain_ a recursive type. The current behavior, which I have maintained, is to treat both `A` and `B` as `SelfRecursive`, and accordingly report errors for both.
Fix `tidy` platform-specific code check
I noticed new platform-specific code was introduced outside of `std::sys` ([example](https://github.com/rust-lang/rust/blob/master/library/std/src/thread/available_concurrency.rs)), which should have been checked against by `tidy`. Apparently there are 2 problems with the current check implementation:
- It ignores everything after encountering "mod tests", which is often at the very top of a file.
- There was a bug where when checking the byte immediately before a found string, the first byte of the file was checked instead.
I fixed the bug and made excluding tests a bit more robust by instead adding the following rules:
- Files with a path containing either `tests` or `benches` are excluded.
- A `cfg(...)` containing `test` is excluded.
(Tests are excluded because almost all tests have something like `#[cfg(not(target_os = "emscripten"))]` somewhere.)
The fixed check found some more cases of platform-specific code; for now I have explicitly excluded them and added a FIXME stating that the platform-specific code must be moved to `sys`.
Fix source code line number display and make it clickable again
Fixes https://github.com/rust-lang/rust/issues/85119.
I used the same logic we're using for other codeblocks: putting the line number `<span>`s into the `example-wrap` directly and then add `display: inline-flex` on `example-wrap`.
r? `@jsha`
ensure failing promoteds in const/static bodies are handled correctly
`const`/`static` bodies are the one case where we still promote code that might fail to evaluate. Ensure that this is handled correctly; in particular, it must not fail compilation.
`src/test/ui/consts/const-eval/erroneous-const.rs` ensures that when a non-promoted fails to evaluate, we *do* show an error.
r? `@oli-obk`
Make unchecked_{add,sub,mul} inherent methods unstably const
The intrinsics are marked as being stably const (even though they're not stable by nature of being intrinsics), but the currently-unstable inherent versions are not marked as const. This fixes this inconsistency. Split out of #85017,
r? `@oli-obk`
Fix suggestions for missing return type lifetime specifiers
This pull request aims to fix#84592. The issue is that the current code seems to assume that there is only a single relevant span pointing to the missing lifetime, and only looks at the first one:
e5f83d24ae/compiler/rustc_resolve/src/late/lifetimes.rs (L2959)
This is incorrect, though, and leads to incorrect error messages and invalid suggestions. For instance, the example from #84592:
```rust
struct TwoLifetimes<'x, 'y> {
x: &'x (),
y: &'y (),
}
fn two_lifetimes_needed(a: &(), b: &()) -> TwoLifetimes<'_, '_> {
TwoLifetimes { x: &(), y: &() }
}
```
currently leads to:
```
error[E0106]: missing lifetime specifiers
--> src/main.rs:6:57
|
6 | fn two_lifetimes_needed(a: &(), b: &()) -> TwoLifetimes<'_, '_> {
| --- --- ^^ expected 2 lifetime parameters
|
= help: this function's return type contains a borrowed value, but the signature does not say whether it is borrowed from `a` or `b`
help: consider introducing a named lifetime parameter
|
6 | fn two_lifetimes_needed<'a>(a: &'a (), b: &'a ()) -> TwoLifetimes<'_<'a, 'a>, '_> {
| ^^^^ ^^^^^^ ^^^^^^ ^^^^^^^^^^
```
There are two problems:
- The error message is wrong. There is only _one_ lifetime parameter expected at the location pointed to by the error message (and another one at a separate location).
- The suggestion is incorrect and will not lead to correct code.
With the changes in this PR, I get the following output:
```
error[E0106]: missing lifetime specifiers
--> p.rs:6:57
|
6 | fn two_lifetimes_needed(a: &(), b: &()) -> TwoLifetimes<'_, '_> {
| --- --- ^^ ^^ expected named lifetime parameter
| |
| expected named lifetime parameter
|
= help: this function's return type contains a borrowed value, but the signature does not say whether it is borrowed from `a` or `b`
help: consider introducing a named lifetime parameter
|
6 | fn two_lifetimes_needed<'a>(a: &'a (), b: &'a ()) -> TwoLifetimes<'a, 'a> {
| ^^^^ ^^^^^^ ^^^^^^ ^^ ^^
error: aborting due to previous error
For more information about this error, try `rustc --explain E0106`.
```
Mainly, I changed `add_missing_lifetime_specifiers_label()` to receive a _vector_ of spans (and counts) instead of just one, and adjusted its body accordingly.
rustc_session: Move more option building code from the `options!` macro
The moved code doesn't need to be generated by a macro, it can use a regular (generic) function and type aliases instead.
(The refactoring is salvaged from a branch with different now abandoned work.)
Add primary marker on codegen unit and generate main wrapper on primary codegen.
This is the codegen part of changes extracted from #84062.
This add a marker called `primary` on each codegen units, where exactly one codegen unit will be `primary = true` at a time. This specific codegen unit will take charge of generating `main` wrapper when `main` is imported from a foreign crate after the implementation of RFC 1260.
cc #28937
I'm not sure who should i ask for review for codegen changes, so feel free to reassign.
r? `@nagisa`
Improve support for NewPM
This adds various missing bits of support for NewPM and allows us to successfully run stage 2 tests with NewPM enabled.
This does not yet enable NewPM by default, as there are still known issue on LLVM 12 (such as a weak fat LTO pipeline). The plan is to make the switch after we update to LLVM 13.
Remove `FakeDefId::expect_local()`
This function returned a fake `DefIndex`, with no indication that it was
fake, when it was provided with a `FakeDefId::Fake`. Every use of the
function uses the returned `DefIndex` in a call to
`tcx.local_def_id_to_hir_id()`, which I'm pretty sure would panic if it
were given a fake `DefIndex`.
I removed the function and replaced all calls to it with a call to
`expect_real()` followed by `DefId::expect_local()` (that's a function
on the *real* `DefId`).
Now, in the case that the function is not inlined, the panic location
will be the caller's location, which is more helpful since the panic is
not `expect_real()`'s fault.
This function returned a fake `DefIndex`, with no indication that it was
fake, when it was provided with a `FakeDefId::Fake`. Every use of the
function uses the returned `DefIndex` in a call to
`tcx.local_def_id_to_hir_id()`, which I'm pretty sure would panic if it
were given a fake `DefIndex`.
I removed the function and replaced all calls to it with a call to
`expect_real()` followed by `DefId::expect_local()` (that's a function
on the *real* `DefId`).
Bump stdarch submodule
Major changes:
- More AVX-512 intrinsics.
- More ARM & AArch64 NEON intrinsics.
- Updated unstable WASM intrinsics to latest draft standards.
- Intrinsics that previously used `#[rustc_args_required_const]` now use const generics. See #83167 for more details.
- `std_detect` is now a separate crate instead of a submodule of `std`.
